/* Unload devices */
static int iboffload_release_devices(void)
{
int i;
mca_bcol_iboffload_device_t *device = NULL;
mca_bcol_iboffload_component_t *cm = &mca_bcol_iboffload_component;
opal_pointer_array_t *devs = &cm->devices;
IBOFFLOAD_VERBOSE(10, ("Destroy all devices.\n"));
for (i = 0; i < cm->num_devs; i++) {
device = opal_pointer_array_get_item(devs, i);
IBOFFLOAD_VERBOSE(10, ("Device %s with index %d will be destroyed.\n",
ibv_get_device_name(device->dev.ib_dev), i));
if (NULL != device) {
OBJ_RELEASE(device);
}
}
IBOFFLOAD_VERBOSE(10, ("All devices were destroyed.\n"));
opal_pointer_array_remove_all(devs);
OBJ_DESTRUCT(devs);
/* release device list */
/*ibv_free_device_list_compat(cm->ib_devs);*/
ompi_ibv_free_device_list(cm->ib_devs);
cm->ib_devs = NULL;
IBOFFLOAD_VERBOSE(10, ("All devices destroyed.\n"));
return OMPI_SUCCESS;
}
int orte_util_decode_pidmap(opal_byte_object_t *bo)
{
orte_jobid_t jobid;
orte_vpid_t i, num_procs;
orte_pmap_t *pmap;
int32_t *nodes;
orte_local_rank_t *local_rank;
orte_node_rank_t *node_rank;
orte_std_cntr_t n;
opal_buffer_t buf;
orte_jmap_t *jmap;
bool already_present;
int j;
int rc;
/* xfer the byte object to a buffer for unpacking */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
if (ORTE_SUCCESS != (rc = opal_dss.load(&buf, bo->bytes, bo->size))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
n = 1;
/* cycle through the buffer */
while (ORTE_SUCCESS == (rc = opal_dss.unpack(&buf, &jobid, &n, ORTE_JOBID))) {
/* unfortunately, job objects cannot be stored
* by index number as the jobid is a constructed
* value. So we have no choice but to cycle through
* the jobmap pointer array and look for this entry. Since
* jobs are cleaned up as they complete, check the
* entire array
*/
jmap = NULL;
already_present = false;
for (j=0; j < orte_jobmap.size; j++) {
if (NULL == (jmap = (orte_jmap_t*)opal_pointer_array_get_item(&orte_jobmap, j))) {
continue;
}
if (jobid == jmap->job) {
already_present = true;
break;
}
}
/* unpack the number of procs */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, &num_procs, &n, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* allocate memory for the node info */
nodes = (int32_t*)malloc(num_procs * 4);
/* unpack it in one shot */
n=num_procs;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, nodes, &n, OPAL_INT32))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* allocate memory for local ranks */
local_rank = (orte_local_rank_t*)malloc(num_procs*sizeof(orte_local_rank_t));
/* unpack them in one shot */
n=num_procs;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, local_rank, &n, ORTE_LOCAL_RANK))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* allocate memory for node ranks */
node_rank = (orte_node_rank_t*)malloc(num_procs*sizeof(orte_node_rank_t));
/* unpack node ranks in one shot */
n=num_procs;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, node_rank, &n, ORTE_NODE_RANK))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* if we already know about this job, we need to check the data to see
* if something has changed - e.g., a proc that is being restarted somewhere
* other than where it previously was
*/
if (already_present) {
/* we already have the jmap object, so let's refresh its pidmap
* using the new data - start by cleaning out the old array
*/
for (j=0; j < jmap->pmap.size; j++) {
if (NULL == (pmap = (orte_pmap_t*)opal_pointer_array_get_item(&jmap->pmap, j))) {
continue;
}
OBJ_RELEASE(pmap);
}
/* now use the opal function to reset the internal pointers */
opal_pointer_array_remove_all(&jmap->pmap);
/* set the size of the storage so we minimize realloc's */
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_size(&jmap->pmap, num_procs))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* add in the updated array */
for (i=0; i < num_procs; i++) {
pmap = OBJ_NEW(orte_pmap_t);
/* add the pidmap entry at the specific site corresponding
* to the proc's vpid
*/
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_item(&jmap->pmap, i, pmap))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* add/update the data */
pmap->node = nodes[i];
pmap->local_rank = local_rank[i];
pmap->node_rank = node_rank[i];
}
/* update the #procs */
jmap->num_procs = num_procs;
} else {
/* if we don't already have this data, store it
* unfortunately, job objects cannot be stored
* by index number as the jobid is a constructed
* value. So we have to just add it to the end
* of the array
*/
jmap = OBJ_NEW(orte_jmap_t);
jmap->job = jobid;
jmap->num_procs = num_procs;
if (0 > (j = opal_pointer_array_add(&orte_jobmap, jmap))) {
ORTE_ERROR_LOG(j);
rc = j;
goto cleanup;
}
/* allocate memory for the procs array */
opal_pointer_array_set_size(&jmap->pmap, num_procs);
/* xfer the data */
for (i=0; i < num_procs; i++) {
pmap = OBJ_NEW(orte_pmap_t);
pmap->node = nodes[i];
pmap->local_rank = local_rank[i];
pmap->node_rank = node_rank[i];
/* add the pidmap entry at the specific site corresponding
* to the proc's vpid
*/
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_item(&jmap->pmap, i, pmap))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
}
}
/* release data */
free(nodes);
free(local_rank);
free(node_rank);
/* setup for next cycle */
n = 1;
}
if (ORTE_ERR_UNPACK_READ_PAST_END_OF_BUFFER == OPAL_SOS_GET_ERROR_CODE(rc)) {
rc = ORTE_SUCCESS;
}
cleanup:
OBJ_DESTRUCT(&buf);
return rc;
}
int orte_util_decode_nodemap(opal_byte_object_t *bo)
{
int n;
int32_t num_nodes, i, num_daemons;
orte_nid_t *node;
orte_vpid_t *vpids;
orte_nid_t *nd, *ndptr;
opal_buffer_t buf;
int rc;
uint8_t *oversub;
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s decode:nidmap decoding nodemap",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* if there are any entries already in the node array, clear it out */
if (0 < orte_nidmap.size) {
/* unfortunately, the opal function "remove_all" doesn't release
* the memory pointed to by the elements in the array, so we need
* to release those first
*/
for (i=0; i < orte_nidmap.size; i++) {
if (NULL != (ndptr = (orte_nid_t*)opal_pointer_array_get_item(&orte_nidmap, i))) {
OBJ_RELEASE(ndptr);
}
}
/* now use the opal function to reset the internal pointers */
opal_pointer_array_remove_all(&orte_nidmap);
}
/* xfer the byte object to a buffer for unpacking */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
opal_dss.load(&buf, bo->bytes, bo->size);
/* unpack number of nodes */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, &num_nodes, &n, OPAL_INT32))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s decode:nidmap decoding %d nodes",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), num_nodes));
/* set the size of the nidmap storage so we minimize realloc's */
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_size(&orte_nidmap, num_nodes))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* loop over nodes and unpack the raw nodename */
for (i=0; i < num_nodes; i++) {
node = OBJ_NEW(orte_nid_t);
/* the arch defaults to our arch so that non-hetero
* case will yield correct behavior
*/
opal_pointer_array_set_item(&orte_nidmap, i, node);
/* unpack the node's name */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, &(node->name), &n, OPAL_STRING))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
/* unpack the daemon vpids */
vpids = (orte_vpid_t*)malloc(num_nodes * sizeof(orte_vpid_t));
n=num_nodes;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, vpids, &n, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* unpack the oversubscribed flags */
oversub = (uint8_t*)malloc(num_nodes * sizeof(uint8_t));
n=num_nodes;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&buf, oversub, &n, OPAL_UINT8))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* transfer the data to the nidmap, counting the number of
* daemons in the system
*/
num_daemons = 0;
for (i=0; i < num_nodes; i++) {
if (NULL != (ndptr = (orte_nid_t*)opal_pointer_array_get_item(&orte_nidmap, i))) {
ndptr->daemon = vpids[i];
if (0 == oversub[i]) {
ndptr->oversubscribed = false;
} else {
ndptr->oversubscribed = true;
}
if (ORTE_VPID_INVALID != vpids[i]) {
++num_daemons;
}
}
}
free(vpids);
free(oversub);
/* if we are a daemon or the HNP, update our num_procs */
if (ORTE_PROC_IS_HNP || ORTE_PROC_IS_DAEMON) {
orte_process_info.num_procs = num_daemons;
if (orte_process_info.max_procs < orte_process_info.num_procs) {
orte_process_info.max_procs = orte_process_info.num_procs;
}
}
/* update num_daemons */
orte_process_info.num_daemons = num_daemons;
if (0 < opal_output_get_verbosity(orte_debug_output)) {
for (i=0; i < num_nodes; i++) {
if (NULL == (nd = (orte_nid_t*)opal_pointer_array_get_item(&orte_nidmap, i))) {
continue;
}
opal_output(0, "%s node[%d].name %s daemon %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), i,
(NULL == nd->name) ? "NULL" : nd->name,
ORTE_VPID_PRINT(nd->daemon));
}
}
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}